On Mon, Nov 16, 2026 at 09:28:56AM +0000, [email protected] wrote:
> Thank you for your contribution! Sashiko AI review found 2 potential
> issues to consider:
> - [Medium] drm/i915: unreachable noop check in
>   intel_crtc_complete_noop_async_flip()
> - [Medium] drm/i915: intel_crtc_complete_noop_async_flip() checks
>   cleared uapi.event
> --
> 
> commit 828bac38a449130d5473b2a45f7f646152241cec
> Author: Ville Syrjälä <[email protected]>
> 
> drm/i915: Complete no-op "async flips" by hand
> 
> This commit adds helpers to manually send the vblank event for atomic
> async flip commits when no planes are updated.
> 
> > diff --git a/drivers/gpu/drm/i915/display/intel_display.c 
> > b/drivers/gpu/drm/i915/display/intel_display.c
> > index f79caeb5a82ec..dd12929284bb8 100644
> > --- a/drivers/gpu/drm/i915/display/intel_display.c
> > +++ b/drivers/gpu/drm/i915/display/intel_display.c
> > @@ -1190,6 +1190,42 @@ static void intel_crtc_disable_flip_done(struct 
> > intel_atomic_state *state,
> >     }
> >  }
> >  
> > +static void intel_crtc_complete_noop_async_flip(struct intel_crtc_state 
> > *crtc_state)
> > +{
> > +   struct intel_display *display = to_intel_display(crtc_state);
> > +   struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
> > +
> > +   if (!crtc_state->do_async_flip)
> > +           return;
> > +
> > +   /*
> > +    * If there are planes actually performing the async flip then
> > +    * the commit will be completed via the flip done interrupt
> > +    * handler. For a fully no-op "async flip" the commit must
> > +    * be completed by hand as no interrupt will be raised.
> > +    */
> > +   if (crtc_state->update_planes)
> > +           return;
> 
> [Severity: Medium]
> Can crtc_state->do_async_flip ever be true when crtc_state->update_planes
> is zero?
> 
> intel_crtc_duplicate_state() initializes both fields to zero/false, and the
> only place that sets new_crtc_state->do_async_flip = true is
> intel_plane_atomic_check_with_state() -> intel_plane_atomic_calc_changes()
> for a visible plane, after intel_plane_atomic_check_with_state() has
> already set BIT(plane->id) in new_crtc_state->update_planes:
> 
> drivers/gpu/drm/i915/display/intel_plane.c:intel_plane_atomic_check_with_state()
>  {
>       ...
>       if (new_plane_state->uapi.visible || old_plane_state->uapi.visible)
>               new_crtc_state->update_planes |= BIT(plane->id);
>       ...
>       return intel_plane_atomic_calc_changes(old_crtc_state, new_crtc_state,
>                                              old_plane_state, 
> new_plane_state);
> }
> 
> When no visible planes are in the atomic state, crtc_state->do_async_flip
> remains false (unlike crtc_state->uapi.async_flip), so this function always
> returns early on the first check and never completes a no-op async flip.
> 
> > +
> > +   if (crtc_state->uapi.event) {
> > +           spin_lock_irq(&display->drm->event_lock);
> > +           drm_crtc_send_vblank_event(&crtc->base,
> > +                                      crtc_state->uapi.event);
> > +           spin_unlock_irq(&display->drm->event_lock);
> > +
> > +           crtc_state->uapi.event = NULL;
> > +   }
> > +}
> 
> [Severity: Medium]
> Will crtc_state->uapi.event always be NULL by the time
> intel_crtc_complete_noop_async_flip() runs?
> 
> In intel_atomic_commit_tail(), commit_modeset_enables() runs before
> intel_atomic_complete_noop_async_flips().  When new_crtc_state->do_async_flip
> is true, intel_update_crtc() -> intel_pipe_update_start() calls
> intel_crtc_prepare_vblank_event(new_crtc_state, &crtc->flip_done_event),
> which moves new_crtc_state->uapi.event into crtc->flip_done_event and sets
> new_crtc_state->uapi.event = NULL.  With update_planes == 0, no flip-done
> interrupt will fire to consume crtc->flip_done_event, leaving the event
> unsent and causing drm_atomic_helper_wait_for_flip_done() to time out.

Ack. Need to deal with that and not prepare the flip done
event without a plane to send it.

-- 
Ville Syrjälä
Intel

Reply via email to